Multi-model fusion

Multiple modelscollaborate, one answer

Create custom fusion presets that call multiple AI models in parallel, with a judge model intelligently fusing the best answer. Tap multiple models' intelligence in one request

Create your fusion preset in 3 steps

1

Create a preset on the Console's "Multi-Model Fusion" page

Pick 2-5 models and designate one as the judge

2

Call with {身份码}/{预设名}

Model name
model: abc123/x4k7m
3

The system calls models in parallel; the judge fuses the best answer

The fused result combines each model's strengths for higher-quality answers

How it works

Parallel request → individual answers → judge fusion → best output

01

Parallel calls

On receiving a request, all models in the preset are called simultaneously (up to 3 concurrent), each producing its own answer

02

Collect answers

Aggregates all model answers, preserving each model's strengths and perspective

03

Judge fusion

The judge reviews all answers, extracts the best content, and generates one high-quality fused answer

04

Return result

Returns the fused answer to the user, supporting both streaming and non-streaming modes

Use cases

When should you use multi-model fusion?

Complex reasoning

Math derivation, logical analysis and other hard tasks — let different models try, and the judge picks the best. Higher success than a single model.

Multi-angle answers

Creative writing, brainstorming and other open tasks — different models bring different perspectives, and the fused content is more comprehensive and creative.

Lower error rate

Critical decisions — cross-validate across models; the judge strips out wrong info, greatly reducing the risk of a single model's mistakes.

High-quality translation

Each model translates once, and the judge combines the strengths of each version to produce a translation more faithful, expressive and elegant than any single model.

Parallel & efficient

Models work simultaneously, no serial waiting

Smart judge

AI picks the best automatically, no manual work

Flexible combos

Freely choose combinations of 2-5 models

Protocol-compatible

OpenAI / Anthropic both supported

Billing notes

Each model billed independently

Each participating model is billed at its own route's normal rate, and the judge model is billed separately too. Plan users deduct plan calls.

Private presets

Each user's Fusion presets are fully private and cannot be accessed or called by other users.

FAQ

Q

How many models can I pick at most?

Each Fusion preset supports up to 5 models; the system caps concurrency at 3 (configurable via fusion_max_parallel in the admin backend). Models beyond the concurrency limit are automatically trimmed.

Q

What does the judge model do?

The judge reviews all participating models' answers and synthesizes the best content. We recommend picking a strong reasoning model as judge (e.g. GLM-5.2, Kimi K2).

Q

Is streaming supported?

Yes. Fusion calls the models in parallel internally (forced non-streaming) and outputs the fused answer in standard SSE streaming. Cherry Studio, Cursor and other clients can receive it normally.

Q

Can other users use my Fusion preset?

No. Fusion presets are fully private; only the creator can call them via {your_identity}/{preset_name}.

Q

What's the identity code in the call name?

The identity code is your account's unique identifier (User.InviteCode). You can copy the full call name from the Console's "Multi-Model Fusion" page.

Create your first fusion preset

Go to the Console's "Multi-Model Fusion" page to enable multi-model collaboration

Go create